Enhanced Salient Object Detection in Remote Sensing Images Using Intra-Class and Inter-Class Attention with Pyramid Vision Transformers

Salient object detection (SOD) in optical remote sensing images (ORSI) is important for various applications. However, the complex nature of ORSI, with its cluttered and varied backgrounds, makes SOD challenging. Traditional methods using convolutional neural networks (CNNs) or transformer have difficulty capturing both local details and global contexts needed for accurate detection. Recent methods have tried combining CNNs and transformer to take advantage of both local and global features. However, these hybrid models often have trouble combining features in a way that clearly distinguishes important regions from non-important ones in complex scenes. To overcome these challenges, we propose a new hybrid model that combines the global capabilities of Pyramid Vision Transformers (PVTv2) with inter-class and intra-class attention mechanisms. Our approach helps improve feature differentiation and enhances the accuracy of detecting salient regions by better merging global and local features. Experiments on two standard datasets show that our method performs better in distinguishing salient from non-salient regions at various scales and complexities in ORSI.